Posterior Exploration SMC Optimization

• Posterior Exploration SMC (PE-SMC) is a global optimization method that uses Sequential Monte Carlo sampling with an adaptive posterior exploration subroutine.
• It integrates Sequential Importance Sampling, component-wise Metropolis moves, and EM updates to refine the importance sampling density dynamically.
• Empirical benchmarks demonstrate that PE-SMC reliably finds near-optimal solutions in complex, multimodal landscapes, outperforming several traditional methods.

Posterior Exploration SMC (PE-SMC) refers to a global optimization algorithm that leverages the Sequential Monte Carlo (SMC) sampling framework, featuring an embedded "posterior exploration" (PE) subroutine for robust adaptation of the importance sampling (IS) density and annealing schedule. PE-SMC is designed to efficiently locate global optimizers of multi-modal, non-convex objective functions by tracking a sequence of annealed probability density functions that concentrate on the global maxima of the objective as annealing progresses (Liu, 2015).

1. Construction of Annealed Targets

The foundation of PE-SMC begins with a nonnegative objective function $f : X \subset \mathbb{R}^d \to \mathbb{R}_+$, whose global maximizer(s) are sought. A sequence of annealed target densities $\{\pi_t(x)\}$ is defined by

$$\pi_t(x) \propto [f(x)]^{\beta_t}, \quad x \in X, \quad \beta_t > 0, \quad t=1,2,\ldots$$

where $\{\beta_t\}$ is an increasing sequence ("inverse-temperatures"), typically initialized as $\beta_t = t$. As $\beta_t \to \infty$, the mass of $\pi_t$ increasingly concentrates on the set $\{x : f(x) = f^*\}$, where $f^* = \max_{x \in X} f(x)$. The annealing schedule is not fixed in advance but adaptively determined to maintain trackability of successive targets.

2. Sequential Importance Sampling with Resampling

The core sampling mechanism is a Sequential Importance Sampling (SIS) protocol, with periodic resampling steps. Given a current IS density $q_{t-1}(x)$, typically represented as a mixture of multivariate Student's t distributions, the iteration proceeds as follows:

• Propagation: Draw $\{\pi_t(x)\}$0 i.i.d. particles $\{\pi_t(x)\}$1.
• Weight update: Compute importance weights $\{\pi_t(x)\}$2 and normalize $\{\pi_t(x)\}$3.
• Effective Sample Size (ESS): Calculate $\{\pi_t(x)\}$4 and its normalized form $\{\pi_t(x)\}$5.
• Resampling: If necessary, resample particles multinomially according to $\{\pi_t(x)\}$6, resetting all weights to $\{\pi_t(x)\}$7.

In PE-SMC, the resampling and the "move" step are embedded in the posterior exploration subroutine, which refines both the particles and the IS density online.

3. Posterior Exploration Subroutine

The PE subroutine systematically explores the target distribution and adaptively reconstructs the IS density. Given $\{\pi_t(x)\}$8 particles from $\{\pi_t(x)\}$9 and the target $$\pi_t(x) \propto [f(x)]^{\beta_t}, \quad x \in X, \quad \beta_t > 0, \quad t=1,2,\ldots$$0, PE executes four components:

1. Importance Sampling (IS): Compute normalized importance weights as in the main SIS procedure.
2. Component-wise Metropolis Move: For each particle and each coordinate, propose $$\pi_t(x) \propto [f(x)]^{\beta_t}, \quad x \in X, \quad \beta_t > 0, \quad t=1,2,\ldots$$1 (with $$\pi_t(x) \propto [f(x)]^{\beta_t}, \quad x \in X, \quad \beta_t > 0, \quad t=1,2,\ldots$$2), accepting with probability $$\pi_t(x) \propto [f(x)]^{\beta_t}, \quad x \in X, \quad \beta_t > 0, \quad t=1,2,\ldots$$3.
3. EM Update for Mixture Parameters: Update the mixture parameters $$\pi_t(x) \propto [f(x)]^{\beta_t}, \quad x \in X, \quad \beta_t > 0, \quad t=1,2,\ldots$$4 (weight, mean, covariance) using weighted EM, where the weight for each particle is its normalized importance $$\pi_t(x) \propto [f(x)]^{\beta_t}, \quad x \in X, \quad \beta_t > 0, \quad t=1,2,\ldots$$5 and component responsibilities $$\pi_t(x) \propto [f(x)]^{\beta_t}, \quad x \in X, \quad \beta_t > 0, \quad t=1,2,\ldots$$6 are computed as normalized mixture probabilities.
4. Adaptive Mixture Augmentation: While $$\pi_t(x) \propto [f(x)]^{\beta_t}, \quad x \in X, \quad \beta_t > 0, \quad t=1,2,\ldots$$7 (threshold):
   • Identify $$\pi_t(x) \propto [f(x)]^{\beta_t}, \quad x \in X, \quad \beta_t > 0, \quad t=1,2,\ldots$$8 with maximum weight.
   • Add a new component centered at $$\pi_t(x) \propto [f(x)]^{\beta_t}, \quad x \in X, \quad \beta_t > 0, \quad t=1,2,\ldots$$9 with small covariance and a small mixture weight $\{\beta_t\}$0.
   • Draw additional samples from the new component, refit the mixture with EM, prune components with $\{\beta_t\}$1, and recompute $\{\beta_t\}$2.

This iterative refinement guarantees the IS density dynamically adapts to cover important regions of the parameter space as annealing proceeds.

4. Adaptive Annealing Schedule

The annealing parameter $\{\beta_t\}$3 is chosen to control the degradation of the effective sample size. In each iteration, $\{\beta_t\}$4 is set such that

$\{\beta_t\}$5

where $\{\beta_t\}$6 is recalculated using the particles and updated mixture from PE, and $\{\beta_t\}$7 (empirically $\{\beta_t\}$8 provides suitable cooling). This is solved as a one-dimensional root-finding problem, typically via bisection, ensuring neither too abrupt nor too slow annealing transitions.

5. Convergence Properties and Computational Analysis

• Convergence: As $\{\beta_t\}$9, $\beta_t = t$0 converges weakly to a uniform distribution on the global maxima set $\beta_t = t$1; in the case of a unique global maximizer, to a point mass. If $\beta_t = t$2 stays close to $\beta_t = t$3 (a property enforced via PE), the empirical distributions from the particle system converge in total variation, and the mode estimate converges to $\beta_t = t$4.
• Stopping Criterion: Annealing progresses until no further improvement in the best-so-far $\beta_t = t$5 is achieved over several iterations.
• Computational Complexity: The per-iteration costs are dominated by $\beta_t = t$6 evaluations of $\beta_t = t$7, $\beta_t = t$8 for EM updates of an $\beta_t = t$9-component mixture, and $\beta_t \to \infty$0 for component-wise Metropolis updates.

6. Empirical Benchmarking and Comparative Performance

PE-SMC has been empirically compared against Particle Swarm Optimization (PSO), Particle-Filter Optimization (PFO), and Sequential-Monte-Carlo Simulated Annealing (SMC-SA) across 17 benchmark test functions with dimensionalities of 2, 5, 10, and 20 (Liu, 2015).

Algorithm	Rastrigin 20D (mean ± std, $\beta_t \to \infty$1)	# Functions with Best Avg.
PSO	$\beta_t \to \infty$2	-
PFO	$\beta_t \to \infty$3	-
SMC-SA	$\beta_t \to \infty$4	-
PE-SMC	$\beta_t \to \infty$5	10 / 23

PE-SMC achieved the best average-best-value performance in 10 out of 23 benchmarking cases, and was never the worst. On many highly multimodal 2D functions, it recovered the global maximum to within $\beta_t \to \infty$6 or better. Compared against PFO and SMC-SA, PE-SMC showed a large improvement margin and remained competitive with a tuned PSO in low dimensions while demonstrating superior performance in higher dimensions.

7. Synthesis and Theoretical Context

PE-SMC integrates the global exploratory advantages of annealed SMC with online-adaptive, mixture-based IS density design and an ESS-triggered temperature schedule. The combination of IS, Metropolis moves, EM mixture refinement, new component insertion, and ESS-matched annealing yields a black-box optimization framework with demonstrated theoretical convergence and empirical efficacy on standard benchmarks. The approach is characterized by its ability to robustly identify and track regions of high objective value across complex, multimodal landscapes (Liu, 2015).